Method for training defect detector

1. A method for training defect detector comprising: obtaining a first reference image of a first reference object, wherein the first reference object has a defect and the first reference image has a first label indicating the defect; training a reconstruction model according to a second reference image of a second reference object, wherein a defect level of the second reference object is in a tolerable range with an upper limit, and the second reference object is associated with the first reference object; obtaining a target image of a target object, wherein the target object is associated with the first reference object and the second object; generating a second label according to the target image, the reconstruction model and an error calculating procedure, wherein the second label comprises a defect of the target object; and training a defect detector by performing a machine learning algorithm according to the first reference image, the target image, and the second label; wherein the error calculating procedure comprises a first difference algorithm and a second difference algorithm, and generating the second label according to the target image, the reconstruction model and the error calculating procedure comprises: generating a reconstruction image according to the target image and the reconstruction model; generating a first difference image and a second image respectively by performing the first difference algorithm and the second difference algorithm respectively according to the target image and the reconstruction image; and performing a pixel-scale operation according to the first difference image and the second difference image to generate the second label.

2. The method for training defect detector of claim 1 , wherein before training the defect detector by performing the machine learning algorithm according to the first reference image, the target image, and the second label, further comprising: training a neural network model according to the first reference image and the first label, wherein the neural network model comprises a refinement network; and selectively removing the second label by the refinement network according to the target image.

3. The method for training defect detector of claim 2 , wherein the neural network model is Faster R-CNN and the refinement network is a classifier of the neural network model.

4. The method for training defect detector of claim 1 , wherein the first difference algorithm comprises: generating a first feature map according to the target image and a neural network model; generating a second feature map according to the reconstruction image and the neural network model; and calculating a difference degree between the first feature map and the second feature map, wherein the first difference image comprises the difference degree.

5. The method for training defect detector of claim 1 , wherein the second difference algorithm comprises: calculating a relative error according to each pixel of the reconstruction image and the target image.

6. The method for training defect detector of claim 1 , wherein the pixel-scale operation is bitwise multiplication.

7. The method for training defect detector of claim 1 , wherein after training the defect detector by performing the machine learning algorithm according to the first reference image, the target image, and the second label, further comprising: obtaining a testing image, wherein the testing image is an image of a testing object associated with the first reference object; selecting a bounding box from the testing image by the defect detector and outputting a feature parameter associated with the bounding box, wherein the bounding box comprises a defect of the testing object; and performing a classification algorithm according to the bounding box and the feature parameter to determine whether the testing object conforms to a specification.

8. A method for training defect detector comprising: obtaining a first reference image of a first reference object, wherein the first reference object has a defect and the first reference image has a first label indicating the defect; training a defect detector by performing a machine learning algorithm according to the first reference image and the first label; training a reconstruction model according to a second reference image of a second reference object, wherein a defect level of the second reference object is in a tolerable range with an upper limit, and the second reference object is associated with the first reference object; obtaining a target image of a target object, wherein the target object is associated with the first reference object and the second object; generating a second label according to the target image, the reconstruction model and an error calculating procedure, wherein the second label comprises a defect of the target object; and calibrating the defect detector by performing the machine learning algorithm according to the target image and the second label; wherein the error calculating procedure comprises a first difference algorithm and a second difference algorithm, and generating the second label according to the target image, the reconstruction model and the error calculating procedure comprises: generating a reconstruction image according to the target image and the reconstruction model; generating a first difference image and a second image respectively by performing the first difference algorithm and the second difference algorithm respectively according to the target image and the reconstruction image; and performing a pixel-scale operation according to the first difference image and the second difference image to generate the second label.

9. The method for training defect detector of claim 8 , wherein before calibrating the defect detector by performing the machine learning algorithm according to the target image and the second label, further comprising: training a neural network model according to the first reference image and the first label, wherein the neural network model comprises a refinement network; and selectively removing the second label by the refinement network according to the target image.